[openib-general] Re: Re: Re: Userspace testing results (many kernels, many svn trees)
Michael S. Tsirkin
mst at mellanox.co.il
Tue Jan 24 09:10:11 PST 2006
Quoting r. Grant Grundler <iod00d at hp.com>:
> Subject: Re: [openib-general] Re: Re: Re: Userspace testing results (many?kernels, many svn trees)
>
> On Tue, Jan 24, 2006 at 08:26:25AM -0800, Shirley Ma wrote:
> > Why not removing the platform dependency and using gettimeofday() instead
> > to measure the performance?
>
> The cycle counters can give us transaction level timing.
> We can see how expensive the first/last transactions are
> and what the variation is during the run.
> gettimeofday() can't do that on most platforms.
>
> AFAICT, only PPC has real wierdness with get_cycles().
> And if that can't be fixed, then we will have to use
> gettimeofday() on PPC or make it a compile time option.
>
> thanks,
> grant
>
No, I think I have a sane work-around: run gettimeofday several times
and time it, compare the result with /proc/cpuinfo.
Also good for stuff like power management clocking the CPU down, etc.
--
MST
More information about the general
mailing list